Skip to content

README.md

Latest commit

 

History

History
119 lines (94 loc) · 5.27 KB

README.md

File metadata and controls

119 lines (94 loc) · 5.27 KB
Raw

OpenWork Station

An ambient assistant, distributed as an OpenWork App.

While listening is switched on, Station notices what you are working on, prepares useful context at the edge of your screen, and hands it to OpenWork when you accept it. It never acts on your behalf: it researches, ranks, and drafts, and every outward step stays yours.

Station is also the reference implementation for the OpenWork Apps platform. It is a genuine third-party consumer of the public contract — it imports nothing from OpenWork, uses no monorepo path aliases, and reaches the host only through the versioned capability broker. If Station can be built this way, so can anything else.

Status

Early. This repository currently contains the manifest, the portable core, and its tests. The realtime surface and background runtime are being ported from the original in-tree prototype. Track the platform work in different-ai/openwork#3303.

There is no published release yet, so OpenWork cannot install Station from this repository today. That is the next milestone, not a claim already met.

What Station asks for, and why

Permission Why
audio.microphone Transcribe what you say while listening is on
ai.realtime Stream that speech to a transcription model
ai.inference.transient Decide whether a moment is worth researching
openwork.connect.read Read-only Slack search, Gmail search, and calendar events
runtime.background.continuous Notice things while you work, not only while a window is open
openwork.threads.start Hand a prepared goal to OpenWork when you accept it
openwork.attachments.create Attach the relevant excerpt, if you choose to include it
desktop.floatingSurface Keep a small island at the edge of the screen
desktop.globalShortcut Start and stop listening without leaving your current app
network.host Reach api.openai.com, and nothing else
storage.app Remember dismissed cards and your transcript preference

Station never receives your OpenAI key. OpenWork holds it, mints a short-lived realtime credential, and hands Station that. Station is told only whether the key is configured.

Starting a thread and creating an attachment additionally require a fresh user gesture — a token the host issues on real input and that expires in seconds. Station cannot start a thread on its own, however confident it is.

Privacy

  • Audio is transcribed only while listening is on, and is never written to disk.
  • The transcript lives in memory for the session and is dropped when you stop.
  • The whole transcript is never attached to a thread. You choose whether an excerpt is included, and even then it is a bounded window of the recent tail.
  • Audit records carry what happened — never transcript text, connected-record contents, or secret values. tests/core.test.ts asserts this.

Architecture

src/core/         portable logic, no DOM and no host coupling
  types.ts          Station's own domain types
  capabilities.ts   the slice of the App protocol Station uses
  relevance.ts      decay, contextual boost, dedup by source identity
  history.ts        card navigation and dismissal
  goal.ts           "I'll look into X" proposals and the spoken yes/no
  transcript.ts     reconciling interleaved realtime partials and finals
  scout.ts          the continuous opportunity scout: rate limits, staleness,
                    duplicate suppression, budgets, and the guarantee that
                    "deciding" always ends
  handoff.ts        building the thread payload, and the transcript choice
src/surfaces/     the floating island and card UI
src/background/   the runtime that drives transcription and research

src/core is pure and covered by tests that need no network, no microphone, and no host. The parts that genuinely need those are the parts that get live evidence, not fixtures pretending to be live evidence.

Why the scout is a separate, pure module

Station must not wait for you to stop speaking — waiting for a long silence is what makes an ambient assistant feel dead. So detection runs during speech, on a bounded schedule. That schedule has to be right under pressure: continuous talking, a lost model reply, a workspace switch mid-request, a topic that recurs every few seconds.

Keeping it pure means all of that is testable directly: that a second pass is refused until the interval elapses, that a lost reply cannot wedge the app on "deciding", that a new session inherits none of the previous session's budget, and that stopping invalidates work already in flight so a slow answer can never surface a card after you stopped listening.

Development

pnpm install
pnpm test        # bun test
pnpm typecheck

Validate and package with the OpenWork app tools:

openwork-app validate openwork.app.json
openwork-app pack --root . --out dist/openwork-station-0.1.0.owapp \
  --repository https://github.com/different-ai/openwork-station \
  --tag v0.1.0 --commit "$(git rev-parse HEAD)"

Those tools live in packages/app-tools and are not yet published to npm; until they are, run them from an OpenWork checkout.

License

MIT. See LICENSE.